Artificial Intelligence: a Disruptive Tool for a Smarter Medicine

European Review for Medical and Pharmacological Sciences 2020; 24: 7462-7474 Artificial intelligence: a disruptive tool for a smarter medicine

C.M. GALMARINI, M. LUCIUS

Topazium Artificial Intelligence, Madrid, Spain

Abstract. – OBJECTIVE: Although highly dinary: in just 100 years, life expectancy has successful, the medical R&D model is failing rocketed from approximately 45 to 72 years. at improving people’s health due to a series of However, there is still a lot of work to be done flaws and defects inherent to the model itself. since the figure for all-cause mortality stands at A new collective intelligence, incorporating hu- almost 57 million. This figure includes 47 mil- man and artificial intelligence (AI) could overcome these obstacles. Because AI will play a lions of adults with chronic, non-communicable key role in this new collective intelligence, it diseases, and 6.5 millions of children, of whom is necessary that those involved in healthcare 5.6 million are under the age of five1. In spite of have a general knowledge of how these technol- the massive economic efforts in R&D and inno- ogies work. With this comprehensive review, we vation, global health improvement seems to have intend to provide it. MATERIALS AND METHODS: reached a plateau. The current model of medical A broad-rang- research is showing difficulties due to a series of ing search has been undertaken on institutional 2 and non-institutional websites in order to identi- limitations . Firstly, its current framework mir- fy relevant papers, comments and reports. rors a linear and sequential process, which has RESULTS: We firstly describe the flaws and been the only paradigm accepted so far. In this defects of the current R&D biomedical model linear approach, investment in R&D produces and how the generation of a new collective in- results that are, at best, only proportional to the telligence will result in a better and wiser med- level of investment. Furthermore, in the current icine through a truly personalized and holistic approach. We, then, discuss the new forms of R&D model the outcome information is usually data collection and data processing and the dif- stored in closed and sealed compartments with ferent types of artificial learning and their spe- very limited access to outsiders, sometimes even cific algorithms. Finally, we review the current within companies or institutions. These “Chinese uses and applications of AI in the biomedical walls” not only prevent innovation, they also in- field and how these can be expanded, as well as crease the time and cost of delivering new drugs the limitations and challenges of applying these or medical devices to patients. Additionally, the new technologies in the medical field. CONCLUSIONS: This colossal common effort current model assesses “value” as the amount of based on a new collective intelligence will ex- “assets” an organization owns (patents, chemical ponentially improve the quality of medical re- entities, etc.) and the relevance of an organization search, resulting in a radical change for the bet- in the market is measured based on these assets. ter in the healthcare model. AI, without replac- However, the paucity of data encourages monop- ing us, is here to help us achieve the ambitious olization of resources and discourages the sharing goal set by the WHO in the Alma Ata declaration of 1978: “Health for All”. of information, which in turn hinders innovation. Thus, the current model drives pharmaceutical Key Words: and biotechnology industries to generate innova- Artificial intelligence, Computational intelligence, Computer vision systems, Biomedical research, On- tion and value through the acquisition of assets cology. and discoveries from other stakeholders, which results in overcrowding of companies within this sector. All these hurdles have led to a R&D model Introduction that no longer works as fast as wanted.

Around 300 billion US dollars are invested A New Medical R&D Model Is Needed every year in medical research worldwide. The Investing more financial and human resources result of this collective effort has been extraor- is not enough to modify the current R&D model

7462 Corresponding Author: Carlos M. Galmarini MD, Ph.D; e-mail: [email protected] Artificial intelligence: a disruptive tool for a smarter medicine in a way that would allow to achieve the ultimate “intelligently”4. Top applications in healthcare goal of improving healthcare. A radical change is include drug discovery and development, auto- needed. This should be based on a new form of mated devices, wearables, diagnostic and medical collective intelligence able to overcome the ob- imaging devices, remote monitoring of patients, stacles of the current model. Since it appears that predictive medicine, robotics and management human intelligence is not enough, this new type of healthcare systems6. There are two key factors of collective intelligence should include artificial that allowed progress in the development of AI intelligence (AI)3. We can use this new kind of (Figure 1). Firstly, in recent years, new sources of collective intelligence to create smarter organiza- medical data, such as health monitoring devices, tions, conduct smarter research and solve medical mobile apps or electronic health records (EHRs) problems that, up to now, where thought to have have been developed, leading to the generation no solution4. How and why will AI revolutionize of vast data banks containing patient data. A medical research? It is known that when a process second factor is the access to new analytical tools becomes digitized and powered by information that can mine and process vast amounts of data. flow, its rate of development jumps onto an ex- This allows to perform pattern recognition, data ponential growth path, doubling every two years classification and predictions in a much shorter (Moore’s Law). Indeed, this system encourages in period of time. silico research, which exponentially reduces the research costs and promotes innovation, result- Data Access ing in a wealth of information and a further cost The success of AI relies on its ability to access reduction, thus generating a positive feedback the right kind of data in an appropriate volume. loop where the output dramatically increases, This represents an enormous challenge. It is cal- since all unnecessary processes and expenditure culated that approximately a zettabyte (a trillion on purposes other than the generation of val- gigabytes) of medical data is generated each year. ue have been eliminated. This information-led Due to the existence of different technologies that environment represents a paradigm shift, as it allow to collect data more easily, this amount is promotes knowledge sharing and the discovery of expected to double every two years7,8. innovative solutions. In this new model, the most benefited are those who share more knowledge New Forms of Data Collection and acquire new knowledge faster. Wearables can deliver an unprecedented amount However, AI elicits many fears, as it is thought of data from millions of people. For example, smart- that it has the potential to replace human intelli- phones are one of the most powerful devices for gence. Yet these fears are unfounded, for AI should health monitoring. Indeed, these regular devices be viewed as a tool rather than a replacement. contain sensors, as well as processing and commu- Therefore, what can AI add to human intelligence? nication capabilities that allow data to be collected, Nowadays, we are overwhelmed with information analysed, and shared. Collecting daily measure- from different new sources (e.g., mobile phone ments with a smartphone may reveal nuances that technology, wearable devices, etc.). AI can integrate are not evident in monthly clinic visits. Similarly, these vast amounts of data and identify relevant pat- in recent years, increased attention has been giv- terns in a way that a human mind would be unable en to patient-reported outcomes (PROs), which to do. On the other hand, AI cannot replace human provide a direct indicator of a patient’s health judgment and wisdom. Therefore, human and ma- condition without correction or interpretation by chine intelligence can complement each other5. Be- a clinician9-12. These data are complementary cause AI will play a key role in medical research, it to those collected by the doctor as they provide is necessary that those involved in this field have a additional information on other aspects that are general knowledge of how these technologies work. important to the patient and that, perhaps, clini- Thus, the aim of this review is to offer a comprehen- cians do not expect or take into account. PROs sive background on AI, as well as its current uses thus are essential components in high-quality, and applications in the medical field, and how these patient-centred care13,14. Collection and integra- can be expanded in the future. tion of ePRO into routine care is feasible, as demonstrated by different studies showing how it Almost Human: Artificial Intelligence increases clinician awareness, resulting in an ear- AI is the field of computer science dealing ly response to patient symptoms and ultimately, a with the development of computers that can act better quality of life14,15.

7463 C.M. Galmarini, M. Lucius

Figure 1. Overview of the AI workflow. The first step is data collection. Data can be collected from different devices and databases generated by universities, hospitals, and biotech and pharmaceutical companies. The second step comprises data processing and data modelling, including feature engineering, model training, evaluation and validation. During this stage, the system can process vast amounts of data and deliver an actionable output that can be a pattern, a classification and/or a prediction.

Data Processing al hurdle: patient data must be anonymized and The completeness and quality of the collected pooled. New developments, such as blockchain data will determine the actual usefulness of the could be used to build data systems that allow database, since the performance of any model re- people to easily and securely share their health lies on the data used to train it16. Thus, formatting, data with researchers, while retaining control filtering, and normalizing the collected data is a over the information supplied22. A second im- first critical step, and represent most of the effort in portant step when preparing data for an AI model building any analytical model17,18. Many different is transforming the input data into a language challenges arise when building a useful data- that a computer can understand (a process known base. A first problem is that the inputted dataset as encoding). Encoding must be done in such a may be incomplete. The options for dealing with way that the semantic characteristics of the data missing data include directly inferring the miss- are captured. Usually, data and categorical and ing values or simply removing sparse features. scalar values are represented as binary vectors. Another challenge is that there can be thousands Numerical features are typically zero-centered of potential predictor features in EHRs19. Tra- by a subtraction of the mean value from every ditionally, this problem was tackled by simply data point or normalized to variance. Finally, log considering only a very limited number of com- transformations are used when extreme values monly collected variables20. However, this poses are altering a feature distribution. a problem, since the resulting models may lead Over the last decade, we have witnessed a dra- to inaccurate predictions21. Moreover, not every matic increase in the number of large, highly com- input feature in a given dataset will be relevant plex databases being generated from wide-rang- for predicting the output label. In fact, including ing datasets, including human genome sequencing, irrelevant input features can compromise the per- gene expression profiles, proteomics, metabolomics, formance of the model derived by AI. A process microbiome, high-resolution imaging, EHR, etc.23. called feature selection is often used to identify Within a few years, we will have vast compre- relevant input features. An example of a feature hensive information from millions of patients24. selection technique is to correlate all input fea- One key challenge would be encoding and com- tures with the output labels: only those features bining this diversity of data, stored in multiple that meet a pre-defined threshold of relevance are sites across many different organizations within kept. However, in healthcare there is an addition- the health sector. Platforms are interfaces that allow

7464 Artificial intelligence: a disruptive tool for a smarter medicine people to connect and share data and insights rap- combine multiple trees into one predictive mod- idly, simply and securely. In the healthcare arena, el to improve performance. Furthermore, ker- these platforms can help to connect very unrelated nel methods are a class of algorithms, amongst stakeholders (e.g., patients, practitioners, research- which support vector machine and kernel ridge ers and regulators), who can combine capabilities regression are the most familiar30. The name and share data to overcome the limitations of the ‘kernel’ is due to the use of a kernel or similarity current sequential approach. AI and data analytics function, that maps non-linear input data into a combined have the potential to become “platform higher-dimensional space allowing its processing aggregators”. by linear classifiers. Finally, and differently from the other machine learning methods that focus Algorithms mostly on pattern recognition, reinforced learn- AI involves different applications of advanced ing is oriented to experience-driven sequential computer intelligence. Machine learning is a sub- decision-making. These algorithmic frameworks field of AI that includes a number of data - ana are capable of taking actions in a particular envi- lytics techniques aimed at building predictive ronment to maximize cumulative reward. Since models from multidimensional datasets; in other output labels can be continuous or categorical, words, computers identify patterns from inputted many machine-learning tasks include regression data without being programmed16. The canonical or classification, respectively, where a regression machine learning workflow involves processing task consists in the prediction of continuous out- the input data, training the underlying model put variables and a classification task consists in (which consists of a set of mathematical formulas the prediction of categorical output variables. and statistical assumptions defined by the learned Once the algorithm or set of algorithms have rules), and then, using it to predict new data. The been selected, the trial model must be judged to data inputted to a machine-learning algorithm allow its optimization. The learning process itself typically consist of “features” and “labels” across consists in finding the most relevant parameters a set of samples. Features are the measurements for the model, i.e., those that make it possible or characteristics, either raw or mathematically to translate input data into accurate output data. transformed, are used to classify the samples, These parameters are identified through a series while labels are the outputs that the algorithmic of back-and-forth iterations, where parameters framework intends to predict. Features and la- are tested, the performance of the model is eval- bels can be continuous or categorical. Data may uated, errors are identified and corrected, and require a primary pre-processing. During this then, the process is repeated again. This process stage, missing or spurious elements are identi- continues until the performance of the model can fied and addressed while the remaining data are no longer be improved, as determined by minimi- transformed into a format more adapted for an zation of the model error. Once the most relevant algorithm. This process is called featurization or parameters have been identified, the system can feature engineering. The more appropriate the be tested by making predictions for new input representation of the input data, the more precise- data. If the model is accurate on both the training ly an algorithm will map it to the output data25. and the test data, then, we can say that the model Depending on the type of data and the question has been properly trained. Machine-learning al- to be answered, a variety of machine learning gorithms can be trained using supervised, unsu- algorithms can be applied. Naive Bayes classi- pervised, semi-supervised or reinforced learning fiers are classification algorithms based on the approaches31,32. Supervised learning is applied Bayes’ theorem that, given a set of existing data, when there are labels available for the input calculates the probability of a hypothesis (or data16. In this case, the training data consist of model) being true26. On the other hand, k-near- sets of input features and associated output la- est-neighbour methods calculate the distances bels that are related. The labels are used to train between the test sample and the samples in the the machine-learning framework through iden- training data27,28. Nearest neighbour methods can tification of patterns that can predict the output be utilized for both classification and regres- labels. The goal of the algorithm is to originate a sion. Another approach is decision trees, which function that, given a specific set of input values, are flowchart-like diagrams used to determine predicts the output values to an acceptable dean outcome29. Decision trees are often used in gree of accuracy. Subsequently, when new input meta-algorithms or ensemble methods, which data become available, the predictions using the

7465 C.M. Galmarini, M. Lucius derived model can be directly made. Supervised and used to refine the derived model, improve learning includes classification algorithms (for its accuracy, and develop new hypotheses. How- categorical output variables; e.g., support vector ever, the reliability of machine-learning meth- machine) or regression algorithms (for continuous ods can be compromised by various elements27. output variables; e.g., linear and logistic regres- Excessive variance (“overfitting”) or excessive sion). By contrast, unsupervised approaches are bias (“underfitting”) is always associated to a utilized when the output labels for the input data deficient performance of a model. Overfitting ap- are not known; these methods learn only from pears when the parameters of the training data the patterns in the input data, i.e., they identi- are fit so specifically to the model that they do fy factors about input features without known not have predictive power outside this particular output labels. Thus, the goal of unsupervised dataset; typically, this occurs when the number of approaches is to cluster subsets of the data based parameters is important, and the model becomes on similar features and to identify the number of too complex. The issue of overfitting can be ad- clusters present in the data. Unsupervised learn- dressed by expanding the size of the training set ing cannot produce an independent predictive or by decreasing the model complexity. On the model. Commonly used unsupervised methods other hand, underfitting occurs when the model include clustering algorithms (e.g., hierarchical cannot detect an accurate relationship between clustering) and dimensionality reduction algo- input and output data, or when the available da- rithms (e.g., principal component analysis). Un- ta are inadequate to permit an accurate pattern supervised methods can be advantageous when identification. Underfitting can be addressed by output labels are missing or incorrect, as it can increasing the model’s complexity33. The main still identify patterns, since the clustering is per- way to test the accuracy of a machine-learning formed only on the input data. Of note, the output model is to check whether it can successfully pre- obtained by models built through unsupervised dict unknown output data. This is usually done approaches can be used as input data for models by setting aside a randomly selected portion of built by supervised approaches. This is known as the dataset during the model’s training. When the semi-supervised learning, which may be useful if training is finished, this set-aside data (also called there is a large amount of input data without the validation dataset) are inputted to the model. The corresponding labels. The algorithm learns from degree to which the output data in the validation labelled data to make a prediction for unlabelled set is accurately predicted by the derived model data and identify patterns. Finally, reinforcement provides a measure of its quality. learning involves a set of dynamic algorithmic frameworks that continuously learns from the en- Deep Neural Networks vironment in an iterative fashion using a system Deep Neural Networks (DNNs) is one of the of reward and punishment. This type of learning principal algorithmic frameworks in machine is limited to a particular context because what learning. These are algorithms based on inter- may produce a maximum reward in one situation connected artificial neurons that execute trans- may be directly associated with punishment in formations on non-linear data to identify relevant another. Algorithms learn to predict the flawless predictive features based on examples34,35. DNNs behaviour within a specific context, to maximize are excellent gears for identifying patterns in its performance and recompense. Thus, learning extremely large and complex datasets. A classical occurs via a reward feedback, known as rein- DNN architecture include a number of layers. forcement signal. The first ones mine relevant features from the in- A machine-learning method is selected de- put data, and the last ones correlate the extracted pending on the task, the characteristics of the da- features to outputs. The width of a DNN relates ta, and the labelled or unlabelled nature of the da- to the maximum number of artificial neurons in a ta. As stated above, if the data are labelled, then, layer while its depth, to the number of hidden lay- a supervised method can be applied to generate a ers it comprises. A neuron (also known as a node predictive model by either regression or classifi- or perceptron) is a computational unit with one or cation of the data. If the data are unlabelled, then, more connections receiving the inputs, a transfer learning must be conducted by an unsupervised function associating the inputs, and an output approach. After building a model by the appropri- connection. The neurons in the input layer take ate machine learning method, the outputs must be the raw data and transfer it to the neurons in the validated. New data can be generated, collected hidden layers, where computation happens. The

7466 Artificial intelligence: a disruptive tool for a smarter medicine hidden layers are sequentially connected such defined based on the purpose of the model to be that each of them learns properties about the data trained. Convolutional Neural Networks (CNNs) by taking as input the transformed input or repre- are one of the most common architectures used sentation produced by the previous hidden layer. for deep learning37. They are normally comprised Once this representation is fed to the next layer, of numerous convolutional and pooling layers. this is transformed into a new representation. This structure favours learning abstract features Indeed, in each neuron, input data are combined at increasing scales going from object edges up with a set of coefficients that amplify or quench to entire objects. That is why CNNs are well it so, assigning weights to inputs within the con- suited for image recognition and computer vision text of the model being built. These weighted tasks38,39. CNNs use weight sharing by sliding a inputs are summed, and the result is run through tiny, trainable filter of weights across the input a non-linear activation function, to conclude vector and convoluting each overlapped region whether that signal should progress any further of inputs with the filter. This differs from ful- through the DNN. If the signal is transferred, is ly connected neural networks where activation said that the neuron has been “activated”. Max- units are bound to all inputs of a feature vector. Out, Sigmoid, rectified linear unit (ReLU) or For example, CNN’s “hidden layers” frequently Softmax are some of the most common activation comprise a series of convolution operations that functions. Depending on the activation function extricate feature maps from the input image and being used, the properties of the activation for pooling actions that perform feature aggregation. the DNN can be quite different. Errors originat- The “hidden” layers output is then input into ed from the training labels are back propagated “fully connected” layers that delivers high-level across the DNN and the model tuned to achieve analysis. Finally, the “output” layer produces a higher performance. Learning is the process predictions. Labelled data and thereby supervised of tuning the weights so that the training data learning methods are frequently used to train are represented as exactly as possible. Finally, CNNs end-to-end. the output layer generates a prediction based on There are other commonly used DNN architec- the weighted inputs from hidden layers25. Once tures. Autoencoders are DNNs that learn how to trained, DNNs can be used over new datasets to efficiently compress and encode data, and then, accurately generate output values. DNNs provide reconstruct it into a representation that is as near considerable advances compared to conventional as possible to the original input. These types of machine learning methods. Indeed, DNNs can algorithms have been applied, for example, to be utilized to attack extremely complex prob- the classification of cancer cases using gene ex- lems and identify patterns in datasets that are pression profiles or for predicting the sequence of just too large and complex for a human brain to proteins. Restricted Boltzmann machine learning process. Similar to conventional machine learn- uses a stochastic recurrent neural network com- ing, DNNs can be used with supervised, unsu- posed of two layers, one made of visible units pervised, semi-supervised or reinforced learning and another made of hidden units, with no lateral methods36. connections. This is one of the most relatively DNNs can have a variety of architectures. easy to train DNNs, and allows forming a com- The simplest of all has three layers: an input, a pact and high-level representation of objects36. middle (hidden) and an output (prediction) layer. Restricted Boltzmann machine learning is used Common DNN architectures often incorporate for unsupervised pre-training of deep networks additional functions and connections to the typ- previous to the subsequent training of the super- ical architecture to permit training using deep- vised models, for example, of disordered protein er networks. Nowadays, DNNs can be simply regions or amino acid contacts40,41. Deep Belief build from pre-existing modules, such as Caffe, Networks, used to build probabilistic generative Theano, PyTorch and TensorFlow. These frame- models, are formed by several stacked hidden works differ principally in easiness of use and layers. Each layer is directly connected to the manipulability. One of the principal challenges next layer. Recurrent Neural Networks (RNNs) in deep learning is how to select the proper are unsupervised or supervised deep learning DNN architecture for a given task. The ideal architectures for identifying temporo-sequential architecture should be able to achieve truthful patterns42,43. RNNs can predict the next data point results with the fewest parameters. Therefore, the in a sequence using the previous data. RNNs number and size of the hidden layers needs to be are well suited for analysing protein or DNA se-

7467 C.M. Galmarini, M. Lucius quences, as well as clinical databases. Generative learning algorithms that are capable of shorten- adversarial networks (GANs) seek to synthesize ing time frames for drug discovery and devel- new data indistinguishable from the training data. opment, and enabling to explore the chemical It opposes two neural networks against each oth- space more effectively46-48. DNNs can be used er; one called generator that creates synthesized to improve the prediction of chemical reac- data (e.g., an image), and a discriminator that tions during the synthesis process. Instead, the estimates the probability that a particular sample pharmacological attributes of a molecule can came from the training data rather than from the be estimated using atomistic simulations. The generator44. GANs have, for example, the poten- generation of different molecules with precise tial to speed up preclinical development. Finally, chemical, physical and pharmacological char- streaming algorithms perform on-the-go calcu- acteristics can be done by training GAN net- lations45. Because data are constantly being in- works while reinforced learning can help the de putted at such volume, streaming algorithms try novo design of chemical entities. Also, the es- to record the essence of what it has been inputted tablishment of structure-property relationships while strategically ignoring the rest. Streaming are estimated by machine learning algorithmic algorithms are useful to monitor a database that frameworks49,50. A large number of computa- is constantly being updated. As it happens with tional methods51,52 for in silico drug discovery other machine learning techniques, deep learning and target extension have been established. One models need to be trained and tested on different of these computational approaches is virtual datasets to prevent overfitting and ensure that the screening, that is, the in silico screening of enor- model can be applied to accurately predict new mous chemical libraries to identify small mole- data. cules that bind to a specific protein (drug-target interaction)53. Virtual screening can also detect Applications in Medical Research interactions with other non-specific proteins AI can be applied to improve a varied range of (off-target effects)54. Deep learning also allows medical fields (Figure 2). A key advantage is that to effectively assess compound toxicity based on these new methods can screen volumes of data to its chemical structure55. In biological research, identify patterns in complex biological systems machine learning applications are becoming that otherwise would be missed. ubiquitous, encompassing not only genome annotation, but also interpretation of complex Drug Discovery and Development biological data comprising drug development, The ever-increasing computational power biomarker detection as well as recommendations along with the abundance of extensive datasets on clinical targets56-58. Another interesting ap- have led the scientific community to develop plication is the identification of new indications for existing drugs, a process known as drug repositioning or drug repurposing55,59. Other areas include the prediction of protein binding sites, identification of key transcriptional cancer driv- ers or the prediction of metabolic functions in complex microbial communities60-63.

Computer-Assisted Diagnosis Digitalization of medical images (e.g., ultra- sounds, computed tomography scans, medical pho- tos, etc.) has promoted the development of automated image interpretation, as almost all image-related tasks involve the evaluation of image features, a field where AI excels. Numerous studies have demonstrated the potential to help radiologists when eval- uating a variety of radiological images, including Figure 2. Principal applications of AI in medical research. CT scans for pulmonary nodules and pneumonias, AI can be applied to improve and accelerate both the preclinical and clinical stages of any project, as well as to guide mammograms for breast lesions, and magnetic res- 64-67 and support clinical decisions. NCE: new chemical entity; onance images for brain tumours . AI methods SAR: structure-activity relationship. can also use any type of image data and therefore,

7468 Artificial intelligence: a disruptive tool for a smarter medicine it can be customized to other fields where images can potentially identify patterns that distinguish are used, such as pathology, dermatology and fast-progressing patients from slow-progressing ophthalmology. In pathology, AI is emerging as ones80. a potent tool to assist pathologists when looking In addition, AI has the potential to transform for microscopic lesions in tissue sections68-70. Dig- recruitment in clinical trials, such as matching ital histopathology images can be used as inputs a patient to the most appropriate trial. Alterna- for deep learning algorithms specifically profi- tively, high-quality data collected in a clinical cient, for example, in the discrimination between trial environment are ideal to train AI systems to normal and abnormal tissues71. Deep-learning accurately predict outcomes. The use of “exter- algorithms can also classify skin lesions with a nal” control arms exploiting real-world evidence level of competence equivalent to that of der- (RWE) stored in vast clinical databases from rou- matologists72-74. Similarly, using retinal images tine clinical practice is of relevance to oncology, (fundus or optical coherence tomography), deep particularly in rare malignancies, where patient learning algorithms were able to discriminate recruitment is quite difficult. In these cases, between macular degeneration and diabetic ret- external control arms can be used as “in silico” inopathy with a performance comparable to that comparators in early-phase clinical trials. RWE of experts75,76. Moreover, deep learning revealed can also be of great value to analyse treatment additional traits in retinal images that allowed efficiency in populations typically excluded from to better predict cardiovascular risk (e.g., high clinical trials (e.g., kidney or hepatic dysfunction, blood pressure, heart attacks, etc.)77,78. These fast, brain metastases, etc.) where the risk-benefit ratio scalable methods can be implemented on mobile is not known. Indication expansion is another devices allowing low-cost universal access to area where RWE may be relevant, as it could important diagnostic care anywhere, positive- provide a means of extracting data on clinical ly affecting primary care practices72. All these outcomes during off-label use, potentially broad- studies show how to use AI to perform simple, ening indications for approved therapies. cost-effective, and widely available studies that could help to identify at-risk patients requiring Healthcare Robotics referral to a consultant. Although sometimes confused, AI and robotics are different from each other. Robotics is a Clinical Applications technology that deals with the design and imple- AI has also several applications at the clini- mentation of robots. The most popular concept cal level. Input data from a healthy cohort, can of robots is that of machines programmed to be used to train deep learning algorithms on perform a specific task over and over without any the fundamental characteristics of healthy states. “intelligence”. However, when combined with Subsequently, this algorithm could be fed by AI, the result is an artificially intelligent or smart data from a patient cohort and used to predict robot that can be trained to carry out complex differences between healthy and disease states, tasks that require more thought and adaptation. In identifying differences in regulatory interactions smart robots, AI acts as the brain and the robotics and biomolecules that could be validated and acts as the body. Indeed, AI programs use data explored further. Thanks to AI, physicians would from the real world acquired through robotics to also monitor their patients and check if they are improve their performance; ergo, they help the likely to present a subsequent disease based on robot strategize about future movements based on their clinical status, and potentially prevent it its prior motion. By doing so, the robot is learning through a data-driven selection of interventions79. how to interpret its own actions. Thus, automa- Likewise, AI algorithms could recognize patients tion, another subfield of AI, is concerned with at a high risk of developing certain clinical dis- how to train a robot to interact with the world orders and send an alert to the patient’s doctor. around it in generalized and predictable ways, AI also has the potential to identify non-evident, manipulate objects in interactive environments, clinically relevant patterns and advice for timely and interact with people. Reinforced learning, interventions. This can also contribute to a sig- which obviates the requirement of labelled data, nificant alleviation of side-effects, with less need is used to safely explore a policy space without for additional medications and less admissions to committing errors that harm the system itself or hospital. An overall change in clinical practice others. Advances in reliable machine perception, can also be induced by AI as these technologies including computer vision, force, and tactile sens-

7469 C.M. Galmarini, M. Lucius ing, much of which is driven by machine learn- niably, data and algorithms can be both flawed. ing, will continue to be key enablers to advance Data can certainly be skewed because of the in- the capabilities of robotics. Indeed, robotics can completeness and inaccuracy of existing datasets. potentially enhance human abilities in healthcare Algorithms can be skewed, deliberately or not in in surprising ways. “Surgical” robots are, for ways that seek to achieve particular outcomes. example, capable of performing complex surgi- Machines that learn on biased data will make cal interventions, including minimally invasive biased decisions, posing new ethical problems and “surgeonless” surgical procedures. The Da and raising new questions. For example, how to Vinci system is the standard of care in multiple prevent certain individuals or groups of people laparoscopic practices and is used in nearly three from using flawed automated systems to exploit quarters of a million procedures a year81,82. Ro- them for their own benefit instead of maximiz- botic surgery application strengthens and allows ing their value for the public good? Ethics are a higher productivity, accuracy, and efficiency of the moral principles that govern an individual’s surgical procedures, favouring a faster recovery behaviour or activity. The medical community and improving patient outcomes. must agree on appropriate ethical principles and frameworks to respond to these new challenges. Virtual Medicine Those frameworks are valuable tools as they set The human body is one of the biggest data a basis that has moral authority to build glob- platforms. Computational models can be used as al standards which reduce the risks that come virtual substitutes or representations of a subset with new technologies. Another weakness is the of the patient’s body or the whole (e.g. genetics, necessity of enormous datasets that may not cell or tissue data, etc.)83,84. These virtual repre- be readily accessible37. As described above, AI sentations can integrate data from EHRs, tele- frameworks built on inaccurately or insufficiently medicine, wearables, and other platforms, tools, large training datasets can result in erroneous es- or media85. Available information should include timations and perpetuate biases. Thus, massively past medical history, genetics, environment, diet, large datasets are essential for building accurate lifestyle, behaviours, preferences, socioeconom- models. However, the majority of datasets are ics, location, measurements from wearable tech- orders of magnitude too small for deep learning nologies, and access and adherence to treatment, algorithms. It should also be stressed that AI among others. MRI and CT scans can also be is different from human intelligence in diverse used to produce a virtual physiologic represen- ways; outperforming in one assignment does tation of the patient, reproducing the anatomy, not necessarily suggest excelling in others. One tissue density, organ architecture and dynamic of the biggest challenges is related to regulatory physiology of an individual patient. These virtual bodies, which have started to unlock the gate to representations of a patient can then be used to machine-learning algorithms. To effectively ap- develop a true personalized medicine for each ply AI platforms in the wider clinical care, regu- individual patient19,86-88. In this sense, different latory bodies request that AI applications achieve systems are being created to represent clinical da- results at least as well as experienced physicians. ta in a reliable, ordered, and expandable formats, As the field matures, expectancies for AI accura- such as FHIR (Fast Healthcare Interoperability cy will unsurprisingly become higher and higher. Resources), thereby simplifying data exchange across sites89. These virtual patients will therefore assist as a tool to support clinical decisions. Conclusions

Limitations and Challenges Although highly successful in the past, the AI presents different challenges and limita- current medical R&D model is failing at im- tions that must be addressed in the future. One of proving health due to a series of flaws and the major limitations lies in the fact that machine defects inherent to the model itself. Therefore, or deep learning systems can be considered as we urgently need to generate a new collective “black boxes” focused on predicting outputs from intelligence, combining human and AI, to dra- data, without taking into account the reasoning matically change the current model. Applying or justification by which a particular outcome a global medical approach to the research flow, is obtained90. Because of this, AI is not immune while incorporating the appropriate AI tools, will to the “garbage in, garbage out” problem. Unde- result in better disease prevention, earlier diagno-

7470 Artificial intelligence: a disruptive tool for a smarter medicine sis and improved, cost-effective drug discovery, 3) Galmarini CM. Smarter Medical Research enhancing patient care standards through a truly Through a New Collective Intelligence. Available personalized and holistic approach. Moreover, in at: https://www.bbvaopenmind.com/en/technology/artificial-intelligence/from-hippocrates-to-arti- an information-based environment, such as the ficial-intelligence-moving-towards-a-collective-in- current world, pharma and biotechnology com- telligence/. panies, scientists and physicians will be able to 4) Malone T. Superminds: the surprising power of share their skills and discoveries, contributing people and computers thinking together. Hachette to the advancement of medical research. Even Book Group 2018. patients will have the chance of contributing to 5) Diprose W, Buist N. Artificial intelligence in medi- medical progress by voluntarily sharing their cine: humans need not apply? N Z Med J 2016; own anonymized medical records. Patients may 129: 73-76. stop being spectators and become main actors in 6) Hosny A, Parmar C, Quackenbush J, Schwartz LH, Aerts H the health arena. Thus, a “system-centric” model . Artificial intelligence in radiology. Nat Rev Cancer 2018; 18: 500-510. could become a “patient-centric model”. Finally, 7) Rodriguez MA, Kotagiri R, Buyya R. Detecting per- thanks to the new wealth of information and formance anomalies in scientific workflows us- knowledge that will become available through ing hierarchical temporal memory. Future Gener this new collective intelligence, health policies Comput Syst 2018; 88: 624-635. would change, emphasizing prevention over 8) savage N. The measure of a man. Cell 2017; 169: treatment and promoting the democratization of 1159-1161. diagnosis, which will allow to detect health is- 9) Di Maio M, Basch E, Bryce J, Perrone F. Patient-re- sues much earlier and therefore, treated at much ported outcomes in the evaluation of toxicity of lower costs with targeted drugs. The benefit could anticancer treatments. Nat Rev Clin Oncol 2016; be greater in peripheral areas or in developing 13: 319-325. FDA U countries, without access to care centres of me- 10) . Guidance for industry: patient-reported outcome measures: use in medical product de- dium/high complexity, as primary care doctors velopment to support labeling claims: draft guid- will be able to make difficult diagnoses earlier ance. Health Qual Life Outcomes 2006; 4: 79. and determine which patients should be referred 11) Howie L, Hirsch B, Locklear T, Abernethy AP. As- to other centres to receive specialized care. In sessing the value of patient-generated data to summary, this colossal common effort based on comparative effectiveness research. Health Aff a new collective intelligence will exponentially (Millwood) 2014; 33: 1220-1228. improve the quality of medical research, resulting 12) LeBlanc TW, Abernethy AP. Patient-reported out- in a radical change for the better in the healthcare comes in cancer care – hearing the patient voice at greater volume. Nat Rev Clin Oncol 2017; 14: model. AI, without replacing us, is here to help us 763-772. achieve the ambitious goal set by the WHO in the Fung CH, Hays RD. 91 13) Prospects and challenges in Alma Ata declaration of 1978: “Health for All” . using patient-reported outcomes in clinical practice. Qual Life Res 2008; 17: 1297-1302. 14) Snyder CF, Aaronson NK, Choucair AK, Elliott TE, Conflict of Interest Greenhalgh J, Halyard MY, Hess R, Miller DM, Reeve BB, Santana M Carlos M. Galmarini and Maximiliano Lucius are employ- . Implementing patient-reported ees of Topazium Artificial Intelligence. outcomes assessment in clinical practice: a review of the options and considerations. Qual Life Res 2012; 21: 1305-1314. 15) Basch E, Deal AM, Dueck AC, Scher HI, Kris MG, Hu- Authors’ Contribution dis C, Schrag D. Overall Survival results of a trial Carlos M. Galmarini: conceptualization, writing, review- assessing patient-reported outcomes for symp- ing and editing draft preparation, approval of final version. tom monitoring during routine cancer treatment. Maximiliano Lucius: conceptualization, writing, review- JAMA 2017; 318: 197-198. ing and editing draft preparation, approval of final version. 16) Camacho DM, Collins KM, Powers RK, Costello JC, Collins JJ. Next-generation machine learning for biological networks. Cell 2018; 173: 1581-1592. References 17) Lohr S. For Big-Data Scientists, ‘Janitor Work’ Is Key Hurdle to Insights. Available at: https://www. 1) WHO. Available at: www.who.int. nytimes.com/2014/08/18/technology/for-big-data- 2) Galmarini CM. https://www.bbvaopenmind.com/ scientists-hurdle-to-insights-is-janitor-work.html. en/technology/innovation/smarter-medical-re- 18) Press G. https://www.forbes.com/sites/ search-through-new-collective-intelligence/. gilpress/2016/03/23/data-prepara-

7471 C.M. Galmarini, M. Lucius

tion-most-time-consuming-least-enjoyableda- 37) Mamoshina P, Vieira A, Putin E, Zhavoronkov A. Ap- ta-science-task-survey-says/. plications of deep learning in biomedicine. Mol 19) Rajkomar A, Oren E, Chen K, Dai AM, Hajaj N, Pharm 2016; 13: 1445-1454. Hardt M, Liu PJ, Liu X, Marcus J, Sun M, Sundberg P, 38) Ciresan DC, Meier U, Gambardella LM, Schmidhuber Yee H, Zhang K, Zhang Y, Flores G, Duggan GE, Ir- J. Deep, big, simple neural nets for handwritten vine J, Le Q, Litsch K, Mossin A, Tansuwan J, Wang, digit recognition. Neural Comput 2010; 22: 3207- Wexler J, Wilson J, Ludwig D, Volchenboum SL, 3220. Chou K, Pearson M, Madabushi S, Shah NH, Butte 39) Hinton GE, Salakhutdinov RR. Reducing the di- AJ, Howell MD, Cui C, Corrado GS, Dean J. Scal- mensionality of data with neural networks. Sci- able and accurate deep learning with electronic ence 2006; 313: 504-507. health records. NPJ Digit Med 2018; 1: 18. 40) Eickholt J, Cheng J. DNdisorder: predicting protein Goldstein BA, Navar AM, Pencina MJ, Ioannidis JP. 20) disorder using boosting and deep networks. BMC Opportunities and challenges in developing risk Bioinformatics 2013; 14: 88. prediction models with electronic health records data: a systematic review. J Am Med Inform As- 41) Spencer M, Eickholt J, Jianlin C. A deep learning soc 2017; 24: 198-208. network approach to ab initio protein secondary structure prediction. IEEE/ACM Trans Comput Bi- Drew BJ, Harris P, Zegre-Hemsey JK, Mammone T, 21) ol Bioinform 2015; 12: 103-112. Schindler D, Salas-Boni R, Bai Y, Tinoco A, Ding Q, Lipton ZC, Berkowitz J, Elkan C Hu X. Insights into the problem of alarm fatigue 42) . A critical review of with physiologic monitor devices: a comprehen- recurrent neural networks for sequence learning. sive observational study of consecutive intensive arXiv 2015; 1506.00019. care unit patients. PLoS One 2014; 9: e110274. 43) Lipton ZC, Kale DC, Elkan C, Wetzell R. Learning 22) Maxmen A. AI researchers embrace Bitcoin tech- to diagnose with LSTM recurrent neural networks. nology to share medical data. Nature 2018; 555: arXiv 2015; 1511.03677. 293-294. 44) Guimaraes GL, Sanchez-Lengeling B, Outeiral C, Farias PL, AspuruGuzik A 23) Yu MK, Ma J, Fisher J, Kreisberg JF, Raphael BJ, Ideker . Objective-reinforced genera- T. Visible machine learning for biomedicine. Cell tive adversarial networks (ORGAN) for sequence 2018; 173: 1562-1565. generation models. arXiv 2017; 1705.10843. 24) Torkamani A, Andersen KG, Steinhubl SR, Topol EJ. 45) Hartnett K, Doré A. Best-Ever Algorithm Found High-definition medicine. Cell 2017; 170: 828- for Huge Streams of Data. Available at: https:// 843. www.quantamagazine.org/best-ever-algorithm- found-for-huge-streams-of-data-20171024/. 25) Angermueller C, Parnamaa T, Parts L, Stegle O. Deep learning for computational biology. Mol Syst Biol 46) Oprea TIT, A. Target, chemical and bioactivity da- 2016; 12: 878. tabases - integration is key. Drug Discov Today 26) Hand DJ, Yu K. Idiot’s Bayes - not so stupid after Technol 2006; 3: 357-365. all? Int Stat Rev 2001; 69: 385-398 47) Smalley E. AI-powered drug discovery captures 27) Butler KT, Davies DW, Cartwright H, Isayev O, Walsh pharma interest. Nat Biotechnol 2017; 35: 604- A. Machine learning for molecular and materials 605. science. Nature 2018; 559: 547-555. 48) Sterling T, Irwin JJ. ZINC 15--ligand discovery for 28) Shakhnarovich G, Darrell T, Indyk P. Nearest-neigh- everyone. J Chem Inf Model 2015; 55: 2324-2337. bor methods in learning and vision: theory and 49) Hill J, Mulholland G, Persson K, Seshadri R, Wolver- practice MIT Press, 2005. ton C, Meredig B. Materials science with large- 29) Rokach L, Maimon O. Classification Trees. Spring- scale data and informatics: unlocking new oppor- er US, 2010. tunities. MRS Bull 2016; 41: 399-409. Pulido A, Chen L, Kaczorowski T, Holden D, Little 30) Shawe-Taylor J, Cristianini N. Kernel methods for 50) MA, Chong SY, Slater BJ, McMahon DP, Bonillo B, pattern analysis cambridge univ. Press, 2004. Stackhouse CJ, Stephenson A, Kane CM, Clowes R, James G, Witten D, Hastie T, Tibshirani R 31) . An intro- Hasell T, Cooper AI, Day GM. Functional materials duction to statistical learning with applications in discovery using energy-structure-function maps. R. Springer US, 2013. Nature 2017; 543: 657-664. 32) Rencher AC. Methods of multivariate analysis. 51) Aliper A, Plis S, Artemov A, Ulloa A, Mamoshina P, John Wiley & Sons, INC, 2002. Zhavoronkov A. Deep learning applications for 33) Domingos P. A few useful things to know about ma- predicting pharmacological properties of drugs chine learning. Commun ACM 2012; 55: 78-87. and drug repurposing using transcriptomic data. 34) LeCun Y, Bengio Y, Hinton G. Deep learning. Na- Mol Pharm 2016; 13: 2524-2530. ture 2015; 521: 436-444. 52) ragoza M, Hochuli J, Idrobo E, Sunseri J, Koes DR. 35) Webb S. Deep learning for biology. Nature 2018; Protein-ligand scoring with convolutional neural 554: 555-557. networks. J Chem Inf Model 2017; 57: 942-957. 36) Bengio Y, Delalleau O. On the expressive power of 53) Mayr A, Klambauer G, Unterthiner T, Steijaert M, deep architectures. In: lecture notes in computer Wegner JK, Ceulemans H, Clevert DA, Hochreiter science. Springer Berlin Heidelberg, 2011. S. Large-scale comparison of machine learning

7472 Artificial intelligence: a disruptive tool for a smarter medicine

methods for drug target prediction on ChEMBL. 68) Cruz-Roa A, Gilmore H, Basavanhally A, Feldman M, Chem Sci 2018; 9: 5441-5451. Ganesan S, Shih NNC, Tomaszewski J, Gonzalez FA, Madabhushi A 54) Xu K, Ba J, Kiros R, Cho K, Courville A, Salakhutdin- . Accurate and reproducible invasive ov R, Zemel R, Bengio Y. Show, attend and tell: neu- breast cancer detection in whole-slide images: ral image caption generation with visual attention. A Deep Learning approach for quantifying tumor arXiv 2015; 1502.03044. extent. Sci Rep 2017; 7: 46450. 55) Van Noorden R. Software beats animal tests at 69) Gareau DS, Correa da Rosa J, Yagerman S, Carucci predicting toxicity of chemicals. Nature 2018; 559: JA, Gulati N, Hueto F, DeFazio JL, Suarez-Farinas M, 163. Marghoob A, Krueger JG. Digital imaging biomark- 56) Azuaje F. Computational models for predicting ers feed machine learning for melanoma screen- drug responses in cancer research. Brief Bioin- ing. Exp Dermatol 2016; 26: 615-618. form 2017; 18: 820-829. 70) Litjens G, Sanchez CI, Timofeeva N, Hermsen M, Nagtegaal I, Kovacs I, Hulsbergen-van de Kaa C, Bult 57) Costello JC, Heiser LM, Georgii E, Gonen M, Men- P, van Ginneken B, van der Laak J. den MP, Wang NJ, Bansal M, Ammad-ud-din M, Deep learning as Hintsanen P, Khan SA, Mpindi JP, Kallioniemi O, a tool for increased accuracy and efficiency of his- Honkela A, Aittokallio T, Wennerberg K, Collins topathological diagnosis. Sci Rep 2016; 6: 26286. JJ, Gallahan D, Singer D, Saez-Rodriguez J, Kaski S, 71) Ehteshami Bejnordi B, Veta M, Johannes van Diest Gray JW, Stolovitzky G. A community effort to as- P, van Ginneken B, Karssemeijer N, Litjens G, van sess and improve drug sensitivity prediction algo- der Laak J, Hermsen M, Manson QF, Balkenhol M, rithms. Nat Biotechnol 2014; 32: 1202-1212. Geessink O, Stathonikos N, van Dijk MC, Bult P, Be- ca F, Beck AH, Wang D, Khosla A, Gargeya R, Irshad 58) Yip KY, Cheng C, Gerstein M. Machine learning and H, Zhong A, Dou Q, Li Q, Chen H, Lin HJ, Heng PA, genome annotation: a match meant to be? Ge- Hass C, Bruni E, Wong Q, Halici U, Oner MU, Ce- nome Biol 2013; 14: 205. tin-Atalay R, Berseth M, Khvatkov V, Vylegzhanin A, Dudley JT, Deshpande T, Butte AJ 59) . Exploiting Kraus O, Shaban M, Rajpoot N, Awan R, Sirinukun- drug-disease relationships for computational drug wattana K, Qaiser T, Tsang YW, Tellez D, Annuscheit repositioning. Brief Bioinform 2011; 12: 303-311. J, Hufnagl P, Valkonen M, Kartasalo K, Latonen L, 60) Alipanahi B, Delong A, Weirauch MT, Frey BJ. Pre- Ruusuvuori P, Liimatainen K, Albarqouni S, Mungal B, dicting the sequence specificities of DNA- and George A, Demirci S, Navab N, Watanabe S, Seno S, RNA-binding proteins by deep learning. Nat Bio- Takenaka Y, Matsuda H, Ahmady Phoulady H, Kova- technol 2015; 33: 831-838. lev V, Kalinovsky A, Liauchuk V, Bueno G, Fernan- 61) Ballester PJ, Mitchell JB. A machine learning ap- dez-Carrobles MM, Serrano I, Deniz O, Racoceanu D, proach to predicting protein-ligand binding affin- Venancio R. Diagnostic assessment of deep learn- ity with applications to molecular docking. Bioin- ing algorithms for detection of lymph node metas- formatics 2010; 26: 1169-1175. tases in women with breast cancer. JAMA 2017; 62) Carro MS, Lim WK, Alvarez MJ, Bollo RJ, Zhao X, Sny- 318: 2199-2210. der EY, Sulman EP, Anne SL, Doetsch F, Colman H, Las- 72) Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, orella A, Aldape K, Califano A, Iavarone A. The tran- Blau HM, Thrun S. Dermatologist-level classifi- scriptional network for mesenchymal transformation cation of skin cancer with deep neural networks. of brain tumours. Nature 2010; 463: 318-325. Nature 2017; 542: 115-118. 63) Langille MG, Zaneveld J, Caporaso JG, McDonald 73) Leachman SA, Merlino G. Medicine: the final fron- D, Knights D, Reyes JA, Clemente JC, Burkepile DE, tier in cancer diagnosis. Nature 2017; 542: 36-38. Vega Thurber RL, Knight R, Beiko RG, Huttenhower Haenssle HA, Fink C, Schneiderbauer R, Toberer F, C. 74) Predictive functional profiling of microbial com- Buhl T, Blum A, Kalloo A, Hassen ABH, Thomas L, munities using 16S rRNA marker gene sequenc- Enk A, Uhlmann L. Man against machine: diagnoses. Nat Biotechnol 2013; 31: 814-821. tic performance of a deep learning convolutional 64) Abe H, Macmahon H, Shiraishi J, Li Q, Engelmann R, neural network for dermoscopic melanoma rec- Doi K. Computer-aided diagnosis in chest radiol- ognition in comparison to 58 dermatologists. Ann ogy. Semin Ultrasound CT MR 2004; 25: 432-437. Oncol 2018; 29: 1836-1842. 65) Korfiatis P, Kline TL, Coufalova L, Lachance DH, 75) Gulshan V, Peng L, Coram M, Stumpe MC, Wu D, Parney IF, Carter RE, Buckner JC, Erickson BJ. MRI Narayanaswamy A, Venugopalan S, Widner K, Mad- texture features as biomarkers to predict MGMT ams T, Cuadros J, Kim R, Raman R, Nelson PC, Mega methylation status in glioblastomas. Med Phys JL, Webster DR. Development and validation of a 2016; 43: 2835-2844. deep learning algorithm for detection of diabetic 66) Wang J, Yang X, Cai H, Tan W, Jin C, Li L. Discrim- retinopathy in retinal fundus photographs. JAMA ination of breast cancer with microcalcifications 2016; 316: 2402-2410. on mammography by deep learning. Sci Rep 76) Kermany DS, Goldbaum M, Cai W, Valentim CCS, Li- 2017; 6: 27327. ang H, Baxter SL, McKeown A, Yang G, Wu X, Yan 67) Yao J, Dwyer A, Summers RM, Mollura DJ. Comput- F, Dong J, Prasadha MK, Pei J, Ting MYL, Zhu J, Li er-aided diagnosis of pulmonary infections using C, Hewett S, Ziyar I, Shi A, Zhang R, Zheng L, Hou texture analysis and support vector machine clas- R, Shi W, Fu X, Duan Y, Huu VAN, Wen C, Zhang sification. Acad Radiol 2011; 18: 306-314. ED, Zhang CL, Li O, Wang X, Singer MA, Sun X, Xu

7473 C.M. Galmarini, M. Lucius

J, Tafreshi A, Lewis MA, Xia H, Zhang K. Identifying (R) Xi Single Site & vaginal extraction of the spec- medical diagnoses and treatable diseases by im- imen. Surg Oncol 2020; 44; 254-255. age-based deep learning. Cell 2018; 172: 1122- 82) Morelli L, Guadagni S, Caprili G, Di Candio G, Bog- 1131 e1129. gi U, Mosca F. Robotic right colectomy using the 77) Maxmen A. Deep learning sharpens views of cells Da Vinci Single-Site(R) platform: case report. Int and genes. Nature 2018; 553: 9-10. J Med Robot 2013; 9: 258-261. 78) Poplin R, Varadarajan AV, Blumer K, Liu Y, McCon- 83) Brown SA. Building SuperModels: emerging pa- nell MV, Corrado GS, Peng L, Webster DR. Predic- tient avatars for use in precision and systems tion of cardiovascular risk factors from retinal fun- medicine. Front Physiol 2015; 6: 318. dus photographs via deep learning. Nat Biomed 84) Brown SA. Principles for developing patient av- Eng 2018; 2: 158-164. atars in precision and systems medicine. Front 79) Vemulapalli V, Qu J, Garren JM, Rodrigues LO, Genet 2016; 6: 365. Kiebish MA, Sarangarajan R, Narain NR, Akmaev VR. 85) Lippman H. How apps are changing family medi- Non-obvious correlations to disease management cine. J Fam Pract 2013; 62: 362-367. unraveled by Bayesian artificial intelligence analy- ses of CMS data. Artif Intell Med 2016; 74: 1-8. 86) Highnam G, Mittelman D. Personal genomes and 80) Abelson S, Collord G, Ng SWK, Weissbrod O, Men- precision medicine. Genome Biol 2012; 13: 324. delson Cohen N, Niemeyer E, Barda N, Zuzarte PC, 87) Mirnezami R, Nicholson J, Darzi A. Preparing for Heisler L, Sundaravadanam Y, Luben R, Hayat S, Wang precision medicine. N Engl J Med 2012; 366: 489- TT, Zhao Z, Cirlan I, Pugh TJ, Soave D, Ng K, La- 491. timer C, Hardy C, Raine K, Jones D, Hoult D, Brit- 88) Miotto R, Li L, Kidd BA, Dudley JT. Deep patient: an ten A, McPherson JD, Johansson M, Mbabaali F, Ea- unsupervised representation to predict the future gles J, Miller JK, Pasternack D, Timms L, Krzyzanows- of patients from the electronic health records. Sci ki P, Awadalla P, Costa R, Segal E, Bratman SV, Beer Rep 2016; 6: 26094. P, Behjati S, Martincorena I, Wang JCY, Bowles KM, Mandel JC, Kreda DA, Mandl KD, Kohane IS, Ramo- Quiros JR, Karakatsani A, La Vecchia C, Trichopoulou 89) ni RB A, Salamanca-Fernandez E, Huerta JM, Barricarte A, . SMART on FHIR: a standards-based, in- Travis RC, Tumino R, Masala G, Boeing H, Panico S, teroperable apps platform for electronic health Kaaks R, Kramer A, Sieri S, Riboli E, Vineis P, Foll M, records. J Am Med Inform Assoc 2016; 23: 899- McKay J, Polidoro S, Sala N, Khaw KT, Vermeulen R, 908. Campbell PJ, Papaemmanuil E, Minden MD, Tanay A, 90) Kourou K, Exarchos TP, Exarchos KP, Karamouzis Balicer RD, Wareham NJ, Gerstung M, Dick JE, Bren- MV, Fotiadis DI. Machine learning applications in nan P, Vassiliou GS, Shlush LI. Prediction of acute cancer prognosis and prediction. Comput Struct myeloid leukaemia risk in healthy individuals. Na- Biotechnol J 2014; 13: 8-17. ture 2018; 559: 400-404. 91) WHO. Declaration of Alma-Ata. Available at: 81) Kim W, Choi Y, Cho JY, Yoon YS, Han HS. Robot sin- https://www.who.int/publications/almaata_decla- gle incision left lateral sectionectomy via da Vinci- ration_en.pdf.

7474